3. Table of Contents
Volume 2, Issue 1, February 2008.
Pages
1- 16 Inference Networks for Molecular Database Similarity Searching.
Ammar Abdo, Naomie Salim.
International Journal of Biometrics and Bioinformatics, (IJBB), Volume (2) : Issue (1)
4. Ammar Abdo, Naomie Salim
Inference Networks for Molecular Database Similarity Searching
Ammar Abdo* ammar_utm@yahoo.com
Faculty of Computer Science and Information Systems
Universiti Teknologi Malaysia
Johor Bahru, Skudai, 81310, Malysia
*Corresponding author : Tel : +6- 0143123054, +6-07- 5532637, Fax : +6-07-5532210
Naomie Salim naomie@utm.my
Faculty of Computer Science and Information Systems
Universiti Teknologi Malaysia
Johor Bahru, Skudai, 81310, Malysia
Abstract
Molecular similarity searching is a process to find chemical compounds that are
similar to a target compound. The concept of molecular similarity play an
important role in modern computer aided drug design methods, and has been
successfully applied in the optimization of lead series. It is used for chemical
database searching and design of combinatorial libraries. In this paper, we
explore the possibility and effectiveness of using Inference Bayesian network for
similarity searching. The topology of the network represents the dependence
relationships between molecular descriptors and molecules as well as the
quantitative knowledge of probabilities encoding the strength of these
relationships, mined from our compound collection. The retrieve of an active
compound to a given target structure is obtained by means of an inference
process through a network of dependences. The new approach is tested by its
ability to retrieve seven sets of active molecules seeded in the MDDR. Our
empirical results suggest that similarity method based on Bayesian networks
provide a promising and encouraging alternative to existing similarity searching
methods.
Keywords: Bayesian networks, molecular similarity searching, chemical databases, inference network,
drug discovery.
1. INTRODUCTION
The term chemoinformatics was coined only a few years ago, but it rapidly gained widespread
use. Chemoinformatics is the use of informatics methods to solve chemical problem [42].
Chemoinformatics is now being extensively used by pharmaceutical and agrochemical
companies. The pressure to find new active compounds and bring them to market as quickly as
possible has led many pharmaceutical and agrochemical companies to use information
technology in their product discovery and development processes. Database searching can be
divided into three distinct classes of problem: exact-match searching for the database record that
is identical to the query record, partial-match searching for those database records that contain
the query and best-match searching for those database records that are most similar to the query
International Journal of Biometric and Bioinformatics, Volume (2) : Issue (1) 1
5. Ammar Abdo, Naomie Salim
record. In chemoinformatics, the first two classes correspond to structure searching and
substructure searching, respectively. The provision of best-searching facilities for chemical
database is normally referred to as similarity searching, which involves quantifying the similarity
of a target molecule with all others in the chemical database in terms of a chosen descriptor or
set of descriptors. It is used whenever a potential drug compound, a lead, has been found. The
lead can be further optimised by finding similar compounds to it, with the hope that a similar, but
better drug can be synthesised.
The virtual screening (VS) is widely used to enhance the cost-effectiveness of drug-discovery
programmes by ranking database of chemical structures in decreasing probability of activity, this
prioritisation then means that biological testing can be focused on just those few molecules that
have significant a priori probabilities of activity. There are many different ways in which a
database can be prioritized, here we focus on similarity searching methods. Similarity searching
is one of the most widely used VS approaches. The basic idea underlying similarity searching
based VS is a very simple idea that similar property principle states that structurally similar
molecules tend to have similar properties [1]. According to this principle, any molecule that has
not been tested for biological activity but is structurally similar to a target molecule that is exhibit
the interest activity is also expected to be active. Furthermore the molecules will be ranked in
decreasing order, so that first molecule is more expected to be active than others and so on.
One objective of the computational tools which applied in chemoinformatics was to finding leads
early in a drug discovery project. The effectiveness of any similarity method can vary greatly from
one biological activity to another in a way that is difficult to predict. Moreover, any two similarity
methods tend to select different subsets of actives from a database, consequently it is advisable
to use several similarity search methods where possible [2].
In essence, most of the molecular similarity measures used originates from areas outside
chemoinformatics, particularly from text retrieval. Although chemical structures differ greatly from
other entities that are commonly stored in database, some parallels can be drawn between
chemical database searches and searches on words or documents [3]. The many similarities
between information retrieval and chemoinformatics that have already been identified suggest
that chemoinformatics is a domain of which information retrieval researchers should be aware
when considering the applicability of new techniques that they have developed [4]. During last
two decades many researches has been done to develop different textual information retrieval
techniques. Currently, Bayesian network the best approach to managing probability and to solve
the uncertainty problem in textual information retrieval.
2. MOLECULAR SIMILARITY SEARCHING
In similarity searching, a query involves the specification of an entire structure of a molecule. This
specification is in the form of one or more structural descriptors and this is compared with the
corresponding set of descriptors for each molecule in the database [5]. A measure of similarity is
then calculated between the target structure and every database structure. Similarity measures
quantify the relatedness of two molecules with a large number (or one) if their molecular
descriptions are closely related and with a small number (large negative or zero) when their
molecular descriptions are unrelated. The results of the similarity measure will be used to sort the
database structures into the order of decreasing similarity with the target. The resulting ranked list
of structures will then be returned to the user. There is an extensive and continuing debate about
what sorts of measures are most appropriate [46]. The similarity measure based on the number
of substructural fragments common to a pair of molecules and a simple association coefficient are
the most common at least until now [46]. The performance of different similarity coefficients with
regard to their use in molecular similarity searching has earlier been analyzed. Several methods
have been used to further optimise the measures of similarity between molecules, which include
weighting [49], standardisation [47] and data fusion [46, 48]. Probability-based similarity
International Journal of Biometric and Bioinformatics, Volume (2) : Issue (1) 2
6. Ammar Abdo, Naomie Salim
searching [50] has also been developed on top of the industry-standard vector-space models
(VSM).
A common application of similarity searching is in the rational design of new drugs and pesticides
where the nearest neighbours for an initial lead compound are sought in order to find better
compounds. Similarity searching is also used for property prediction purposes [7], where the
properties of an unknown compound are estimated from those of its nearest neighbours.
Underpinning these applications of molecular similarity measure is the similar property principle
[1], which states that structurally similar molecules will exhibit similar physiochemical and
biological properties. Related to the similar property principle is the concept of neighbourhood
behavior [8], which states that compounds within the same neighbourhood or similarity region
have the same activity. Unknown biological or physicochemical properties of a molecule can be
predicted from the properties of molecules that lie within the same neighbourhood region. In lead
finding, selection of compounds whose neighbourhood regions overlap one another should be
avoided. In lead optimisation, if a particular compound is found to be active, compounds that lie in
the same neighbourhood region can be tested to find one with the most optimum activity.
The first reports on similarity searches appeared in the mid-1980s, based on the work carried out
at Lederle Laboratories [7] and Pfizer [9]. In the Lederle study, molecules were represented by
their constituent atom pairs, where an atom pair is a substructural fragment comprising two non-
hydrogen atoms together with number of intervening bonds. The similarity search allowed users
to request either some number of the top-ranked molecules or all those that had a similarity with
the target structure greater than a minimal value. In the Pfizer system, together with a
conventional substructural query, a user can submit a target molecule typical of the type of the
structure that was required. The conventional screen search and atom-by-atom search were used
to identify matches in the substructure searching, after which a similarity measure based on the
screens common to the target and the matches was used to rank the substructure search output.
The subsequent development of a faster, inverted-file-based, nearest neighbour search algorithm
allowed the ranking of the entire database against the target structure in real time, without the
need for the specification of the initial substructural query. Since the Lederle and Pfizer systems,
similarity searching has undergone further development. An example is Hagadone’s work on
substructure similarity searching [10]. Substructure similarity searching is used to identify
molecules containing a substructure similar to a target structure or substructure. Another
extension of similarity search was described by Fisanick et al. [11] on facilities developed for
Chemical Abstracts Service (CAS) Registry File. It focuses on different types of similarity
relationships that can be identified between a structure in the query and a database structure.
This study found that different representations could give different measures of structural
resemblances between compounds, which suggest that a further analysis into a combined
approach could give a more comprehensive similarity measure between them. The use of
similarity calculations between molecules have since been used not only in similarity searching,
but also in applications like compounds selection [12, 13] and molecular diversity analysis [14, 15,
16]. Three principal tools used for the similarity calculations are the representation that is used to
characterize the molecules that are being compared, the weighting scheme that is used to assign
differing degrees of importance to the various components of these representations, and the
coefficient that is used to determine the degree of relatedness between two structural
representations [17].
2.1 Molecular descriptors
Molecular descriptors are vectors of numbers, each of which is based on some pre-defined
attributes. They are generated from a machine-readable structure representation like a 2D
connection table or a set of experimental or calculated 3D co-ordinates. Molecular descriptors
can be classified into 1D descriptors, 2D descriptors and 3D descriptors. 2D descriptors are
based on information derived from the traditional 2D structure diagram. Examples of 2D
descriptors are 2D fingerprint and topological indices, which are our focus as they play a
prominent role in the experimental work of this paper.
International Journal of Biometric and Bioinformatics, Volume (2) : Issue (1) 3
7. Ammar Abdo, Naomie Salim
2D fingerprints are the most commonly used descriptors. These descriptors were initially
developed to provide a fast screening step in substructure search systems in which bit strings are
used to represent molecules. They have also proved very useful for similarity searching. There
are two different types of 2D fingerprints: dictionary-based bit strings and hashed fingerprints. In
dictionary-based bit strings, a molecule is split up into fragments of specific functional groups or
substructures. The fragments used are recorded in a predefined fragment dictionary that specifies
the corresponding bit positions of the fragments in the bit string. Bits either individually or as a
group represent the absence or presence of fragments. Examples of dictionary-based
assignment are the CAS ONLINE Screen Dictionary for substructure searching [18], Barnard
Chemical Information system [19, 20] and MDL MACCS key system [21, 22]. In hashed
fingerprints, all the unique fragments that exist in a molecule are hashed using some hashing
function to fit into the length of the bit string. This approach allows for more generalisations
because it does not depend on a predefined list of structural fragments. The fingerprints
generated are characterised by the nature of the chemical structures in the database rather than
by the fragments in some predefined list. This approach is used in the Daylight Chemical
Information Systems [24] and Tripos systems [23].
Topological indices characterise the bonding pattern of a molecule by a single value integer or
real number, obtained from mathematical algorithms applied to the chemical graph representation
of the molecules. Each index thus contains information not about fragments or some locations on
the molecule, but rather about the molecule as a whole. Simpler descriptors include the number
of atoms and bonds and the number of rotatable bonds.
Similarity measures based on bit strings are currently the most widely used approach for
database searching [25]. One of the principal applications of bit string based searching is in the
selection of compounds for inclusion in biological screening programs. This is largely due to the
low processing requirements needed to calculate the similarities between a target structure and a
large number of structures.
2.2 Weighting schemes
A weighting scheme is used to differentiate between different features in a molecule, based on
how important they are in determining the similarity of that molecule with another molecule.
Certain molecular features can be emphasised by associating higher weights with them when
calculating similarity. Different types of statistical information can be extracted from computerised
representations of molecules to form the basis for a fragment weighting schemes. These are
follows, (a) Fragment Frequency (ff), is the number of occurrence of a particular fragment within a
molecule, with high frequently occurring fragments being given a greater weight than those that
occur less frequently. (b) Inverse Fragment Frequency (iff), is the frequency of the fragment in the
molecule collection, with less frequently occurring fragment being given a greater weight than
those that occur high frequently throughout the molecule collection. (c) Molecule size (mz), is the
number of the fragments assigned to a molecule, with a fragment in small molecule being
assigned a greater weight than the same fragment in a large molecule. One more weighting
scheme can be used whenever we can differentiate between active and inactive molecules within
dataset. Unfortunately, limited studies have been done on the effect of applied weighting
schemes on molecular similarity searching methods. All of the above mentioned considerations
have been used for assigning weights at the National Cancer Institute [26]. Willett and Winterman
have found that giving more weight to fragments that occur more frequently in a molecule did
seem to give good results, but other weighting schemes had little significance [27].
2.3 Similarity Coefficients
Similarity coefficients are used to obtain a numeric quantification to the degree of similarity
between a pair of structures [28]. There are four main types of similarity coefficients [29, 30, 31] :
distance coefficients, association coefficients, correlation coefficients and probabilistic
coefficients. Association coefficients are commonly used with binary representations and are
often normalized to lie within the range of zero (no similar features in common) and unity
(identical representations). However, they can be used with non-binary representations, in which
International Journal of Biometric and Bioinformatics, Volume (2) : Issue (1) 4
8. Ammar Abdo, Naomie Salim
case the range may be different. Correlation coefficients measure the degree of correlation
between sets of values characterizing a pair of objects. Distance coefficients quantify the degree
of dissimilarity between two objects and, when normalized and using binary data, range between
zero (identity) and unity (no similar features in common). Probabilistic coefficients, whilst not
much used in measuring molecular similarity, focus on the distribution of the frequencies of
descriptors over the members of a data set, giving more importance to a match on an infrequently
occurring variable. Examples of these coefficients can be found elsewhere [29]. Assume SK,L is
the similarity between molecules K and L, both molecules described by binary representation. For
bit string descriptors, n is the total bit positions in the bit strings representing the two molecules
compared. b is the number of bit positions set in only one of the two molecules whilst c is the
number of bit positions set in only the other molecule. d of the n bits are not set in either one of
the molecules and a is the number of bits set in both molecules. Thus, n = a + b + c + d. The
origins of the coefficients can be found in a review paper by Ellis et al. [31]. Examples of some of
the coefficients that were used are listed in Table 1.
Continuous Binary
Coefficient
Formula Range Formula Range
(w w ) M
∑
j =1
jk jl
a
∑ (w ) + ∑ (w ) − ∑ (w w jl )
Tanimoto M 2 M 2 M
-0.3 to 1 0 to 1
j =1
jk
j =1
jl
j =1
jk a + b + c
∑
M
j =1
(w jk
w jl
)
a
Cosine
∑
M
j =1
(w ) jk
2
∑
M
j =1
(w ) jl
2 0 to 1
( a + b )( a + c )
0 to 1
M
n ∑ (w jk
w jl
)
j =1
n×a
Forbes M
2
M
2 -∞ to ∞ 0 to ∞
∑ w jk ∑ w jl
( a + b )( a + c )
j =1 j=1
M
∑ w jk w jl
j =1 a
Russell-Rao -∞ to ∞ 0 to 1
n n
2 ∑ (w jk w jl )
M
j =1
2a
Dice M
∑ w jk
j =1
( ) 2
+ ∑ (w jl )
M
j =1
2 0 to 1
2a + b + c
0 to 1
TABLE 1: Examples of Association Coefficients.
Tanimoto coefficient in Eq. 1 is the most popular coefficient used by similarity methods. If two
molecules K and L have b and c bits set in their fragment bit-strings, with a of these bits being set
in both of the fingerprints, then the similarity between these two molecules using Tanimoto
coefficient is defined to be:
a
SK ,L = (1)
a+b+c
The Tanimoto coefficient gives values in the range of zero (no bits in common) to unity (all bits
the same). The Tanimoto coefficient gives the best result than the other coefficients. Currently,
International Journal of Biometric and Bioinformatics, Volume (2) : Issue (1) 5
9. Ammar Abdo, Naomie Salim
The Tanimoto coefficient is widely used in molecular similarity methods and was becomes the
best choice in both in-house and commercial software systems for chemical information
management.
3. BAYESIAN NETWORKS
Recent research in information retrieval has proved that retrieval models based on Bayesian
network give significant improvements in retrieval performance compare to conventional models
[36, 37, 38, 43]. It is therefore likely that Bayesian network is able to represent the main
(in)dependence relationships between molecular descriptors as conditional probabilities with the
degree of resemblance between pairs of such descriptors computed to represent the probability.
Molecular similarity will be regarded as an inference or evidential reasoning process in which the
probability that a given compound met the requirements of a query is estimated and used as
evidence. Network representations have show promise as mechanisms for inferring these kinds
of relationships. In this paper, we explore the possibility and effectiveness of using such
networks for similarity searching.
A Bayesian network (BN) is graphical model of a probability distribution [33]. A Bayesian network
is a directed acyclic graph (DAG) in which the nodes represent random variables and the arcs
show causality, relevance or dependency relationships between them. The variables and their
relationships comprise the qualitative knowledge stored in a Bayesian network. The strength of
the relationships, measured by means of probability distributions, is also stored in the DAG.
Associated with each node is a set of conditional probability distributions, one for each possible
combination of values that its parents can take. A Bayesian network can be considered an
efficient representation of a joint probability distribution that takes into account the set of
independent relationships represented in the graphical component of the model. In general terms,
given a set of variables {X1, . . . , Xn} and a Bayesian network G, the joint probability distribution in
terms of local conditional probabilities is obtained as follows:
n
P ( X 1 ,... X n ) = ∏ P ( X i π ( X i ))
i =1
where π(Xi) is any combination of the values of the parent set of Xi. If Xi has no parents, then the
set π(Xi) is empty, and therefore P(Xi|π(Xi)) is just P(Xi). Once completed, a Bayesian network
can be used to derive the posterior probability distribution of one or more variables in the network,
or to update previous conclusions when new evidence reaches the system.
4. SIMILARITY INFERENCE NETWORK MODEL
The basic model for similarity inference network, shown in Fig.1, consists of two component
networks: a compound network and a query network. The compound network represents the
compound collection. The compound network is built once for a given collection and its structure
does not change during query processing. The query network consists of a single node, which
represents the target molecule and one or several query molecules, which express the target
molecule. A query network is built for each target molecule and modified during query processing
as the query is refined or additional representations are added in an attempt to better
characterize the target molecule. The compound and query networks are connected though links
between their descriptor nodes.
4.1 Compound Network
The compound network shown in Fig. 1 is a simple direct acyclic graph (DAG) consisting of
compound nodes (cj) as roots, and descriptor nodes (di) as leaves. Each compound node
represents a compound in the collection. Each compound node has a prior probability associated
International Journal of Biometric and Bioinformatics, Volume (2) : Issue (1) 6
10. Ammar Abdo, Naomie Salim
with it that describes the probability of observing that compound. This prior probability will
generally be set to 1/(collection size) and this probability will be small for real collections.
Compound nodes have one or more descriptor nodes as children. The descriptor nodes can be
divide into several subsets, each corresponding to a single descriptor technique that has been
applied to the compound. When 1052 bits are used to describe the compounds using BCI
fingerprint, 1052 nodes are used to represent these bits. If 10 topological indices are used to
describe the compounds, 10 nodes are used to represent these numerical values. We represent
the assignment of a specific descriptor to a compound by draw a directed arc to the descriptor
node from each compound node corresponding to a descriptor node. Each descriptor node
contains a specification of the conditional probability associated with the node given its set of
parent compound nodes. This specification incorporates the effect of any weighting scheme
associated with the descriptors node.
C1 C2 Cj CM
d1 d2 d3 di dN
Q
FIGURE 1: Similarity inference network model.
4.2 Query Network
The query network is an “inverted” DAG with a single leaf that corresponds to a target molecule
and multiple roots that correspond to the descriptors that express the target. If there is only one
query molecule, the target molecule node and query molecule node coincide. In addition, the
query network is intended to allow us to combine several query molecules to form a single query
molecule. The roots of the query network are query descriptors; they correspond to the
descriptors used to express the target molecule. A single query descriptor node has a single
compound descriptor node as parent. Each query descriptor node contains a specification of its
dependence on a single parent compound descriptor node. The query descriptor nodes define
the mapping between the descriptor layer used to represent the compound collection and the
descriptor layer used to describe target molecule. In our model, the relation between query and
compound descriptors is 1:1 and completely depends. Thus, in order to simplify and reduce our
model, the query descriptors are the same as the compound descriptors. The attachment of the
query descriptors nodes to the compound network has no effect on the basic structure of the
compound network. None of the existing links needs change and none of the conditional
probability specifications stored in the nodes are modified.
To produce a ranking of the compounds in the collection with respect to a given target molecule
T, we compute the probability that this target molecule is satisfied given that compound cj has
been observed, P(T|cj). This is referred to as instantiating cj and corresponds to attaching
evidence to the network, by stating that cj = true, whereas the rest of the compound nodes are set
to false. When the probability P(T|cj) is computed, this evidence is removed and a new compound
cj, i ≠ j , is instantiated. By repeating this computation for the rest of the compounds in the
collection, the ranking is produced.
International Journal of Biometric and Bioinformatics, Volume (2) : Issue (1) 7
11. Ammar Abdo, Naomie Salim
The similarity inference network is intended to capture all of the significant probabilistic
dependencies among the random variables represented by nodes in the compound and query
networks. If these dependencies are characterised correctly, then the results provided are good
estimates of the probability this target molecule is met. Given the prior probabilities associated
with the compounds (roots) and the conditional probabilities associated with the interior nodes
(descriptor nodes), we can compute the posterior probability associated with each node in the
network. Further, if the value of any variable represented in the network becomes known we can
use the network to recompute the probabilities associated with all remaining nodes based on this
“evidence”. The query network is first built and attached to the compound network, and then the
belief associated with each node in the query network computed. All compounds are equally likely
(or unlikely).
4.3 Probabilities Estimation
For any of the non-root nodes A of the network, the dependency on its set of parent nodes {P1,
P2,…,Pn}, quantified by the conditional probability P(A|P1,P2,..,Pn), must be estimated and
encoded. Link matrices are used to encode the probability value assigned to a node A given any
combination of values of its parent nodes. However, all the random variables (di, q, T),
represented by the non-root nodes in the network, are binary and therefore, when a node has n
n
parents, the link matrix associated with it is of size 2 x 2 .
Canonical link matrix forms allow us to compute for A any value LA[i, j] of its link matrix LA, where
n
i Є {0,1} and 0 ≤ j ≥ 2 , will be used [36, 40]. The row number {0,1} of the link matrix corresponds
to the value assigned to the node A, whereas the binary representation of the column number is
used so that the highest order bit reflects the value of the first parent, the second highest order bit
the value of the second parent and so on. The weighted-sum canonical link matrix form [36]
allows us to assign a weight to the child node A, which is, in essence, the maximum belief that
can be associated with that node. Furthermore, weights are also assigned to its parents,
reflecting their influence on the child node. Consequently, our belief in the node is determined by
the parents that are true. For instance if node A has two nodes as parent P1, P2 and that the
weight assigned to them w1, w2 respectively and wA is weight for node A, now suppose
P(P1=true)=p1 and P(P1=true)=p2, then the link matrix LA is as follows:
w2wA w1wA (w1 + w2 )wA
1 1− w + w 1− w + w 1− w + w (2)
LA = 1 2 1 2 1 2
0 w2wA w1wA (w1 + w2 )wA
w1 + w2 w1 + w2 w1 + w2
The evaluation for this link matrix is as following:
( w1 p 1 + w 2 p 2 ) w A
P ( A = true ) = (3)
w1 + w 2
( w1 p 1 + w 2 p 2 ) w A (4)
P ( A = false ) = 1 −
w1 + w 2
In the more general and complicated case of the node A having n parents, the link matrix at Eq. 2
cannot be evaluated because become NP hard, therefore the derived link matrix can be
evaluated using the following closed form expression:
n
wA ∑ wi pi
i =1
bel ( A) = n
(5)
∑w
i =1
i
International Journal of Biometric and Bioinformatics, Volume (2) : Issue (1) 8
12. Ammar Abdo, Naomie Salim
For our similarity inference network model, estimates for the (dj, q, T) random variables that
characterise the following three dependencies are provided
• The dependence of the descriptor nodes upon the compound nodes which containing them
• The dependence of the query molecule nodes upon the descriptor nodes which containing
them.
• The dependence of the target molecule upon the different query node.
In case one query molecule node is used in the model, then the target molecule node coincide
with query molecule node. Therefore, we only need to estimate the first two probabilities. The
only roots in Fig. 1 are the compound nodes, therefore the prior probability associated with these
nodes is set to 1/(collection size). Compound and query descriptor nodes are viewed as identical
under the assumption that the user knows the set of compound descriptors and can formulate
queries using the compound descriptors directly.
To estimate the probability that a descriptor node is good for discriminating a chemical
compound’s structure, a weighting function can be incorporated in the weighted-sum link matrix.
We will use the weighting schemes mentioned in section 2.2 above and difference between
values of descriptors nodes for compound and query as weighting function. For instance,
molecular descriptors such as topological indices values and bit frequency of fingerprints can be
used for weighting function. For normalized topological indices descriptor, this estimate is given
by:
2
P (d i c j = true) = α + (1 − α ) × (1 − d i − d i' ) (6)
where α is a constant and experiments using the inference network show that the best value for α
is 0.4 [36, 40], di is the value of compound descriptor and dj’ is the value of query descriptor. For
bit string molecular descriptors, the molecule size (mz) and inverse fragment frequency (iff) as
weighting functions. This estimate is given by:
k jq
P ( d i c j = true ) = α + (1 − α ) × × iff i (7)
mz j
For both descriptors,
P(di all parent false) = 0 (8)
Where kjq is the no of common bits between q and cj, mzj is the size of compound cj and iffi is the
inverse fragment frequency of fragment i in the compound collection.
The target molecule can be expressed as a small number of queries. These can be combined
using a weighted-sum link matrix in Eq. 3 with weights adjusted to reflect any user judgments
about the importance or completeness of the individual queries. We only have one query node,
so the wA in probability function in Eq. 5 will omit and wi is set to 1 that’s for topological indices
and incorporated with weighting function given below for bit strings
n k
∑
jq
( × iff i × p i )
i =1 mz q (9)
bel ( Q ) = n k
∑
jq
( × iff i )
i =1 mz q
where kjq is same as in Eq. 7, mzq is the size of query q and iffi is the inverse fragment frequency
of fragment i in the compound collection. The kjq factor is normalizing to the range [0, 1] by
International Journal of Biometric and Bioinformatics, Volume (2) : Issue (1) 9
13. Ammar Abdo, Naomie Salim
dividing kjq by the maximum possible kjq value (mzj and mzq are the maximum values of kjq in Eq.
7 and Eq. 9 respectively). The inverse fragment frequency is given by
collection size
iff = log( ) (10)
fragment frequency
We will normalize iff to the range [0, 1] by dividing iff by the maximum possible iff value in the
collection (the iff score for a fragment that’s occurs once).
collection size
log( )
fragment frequency
iff = (11)
log(collection size)
5. EXPERIMENTAL DESIGN
In this study a subset of the MDDR database comprised of around 15 biologically active groups of
compounds have been used. Most of the activities chosen are highly diverse whereas the first
four categories can be regarded as the most heterogeneous as compared to the rest of the
compounds. The experiments were conducted using a collection of 1360 compounds from the
MDL’s Drug Data Report (MDDR) database [44]. For the first experiment developed to test our
similarity inference model with 2D fingerprint descriptors. We used bit string descriptors from
Barnard Chemical Inc (BCI) fingerprint generation software based on BCI dictionaries bci1052
[41] for 1052 bit-strings. Unfortunately this type of fingerprint only represents the fragment
presence without frequency counts. Therefore, fragment frequency for any fragment in the
compound is set to 1. We used 9 targets molecules as queries for each of the 7 activity groups.
The main groups, their subgroups and their aggregate activity are summarized in Table 2
No.
S.No Activity
Molecules
Interacting on 5HT receptor
5HT Antagonists 48
1 5HT1 agonists 66
5HT1C agonists 57
5HT1D agonists 100
Antidepressants
2 Mao A inhibitors 84
Mao B inhibitors 148
Antiparkinsonians
31
3 Dopamine (D1) agonists
103
Dopamine (D2) agonists
Antiallergic/antiasthmatic
73
4 Adenosine A3 antagonists
150
Leukotine B4 antagonists
Agents for Heart Failure
5
Phosphodiesterase inhibitors 100
AntiArrythmics
6 Potassium channel blockers 100
Calcium channel blockers 100
Antihypertensives
7 ACE inhibitors 100
Adrenergic (alpha 2) blockers 100
Total molecules 1360
TABLE 2: Groups and activities of the dataset.
International Journal of Biometric and Bioinformatics, Volume (2) : Issue (1) 10
14. Ammar Abdo, Naomie Salim
For the second experiment developed to test our similarity inference model with topological
indices, we generated around 100 topological indices using the Dragon software [45], out of
which only 10 have been selected, accounting for around 98% of the variance in the dataset. A
list of the 10 topological indices selected is shown in Table 3. Results were compared with the
industry standard Tanimoto measure [46].
TI Description
Gnar Narumi geometric topological index
Xt Total structure connectivity index
Dz Pogliani index
SMTI Schultz Molecular Topological Index
PW3 path/walk 3 – Randic shape index
PW4 path/walk 4 – Randic shape index
PW5 path/walk 5 – Randic shape index
PJI2 2D Petitjean shape index
CSI eccentric connectivity index
D/Dr03 distance/detour ring index of order
TABLE 3: Selected Topological Indices.
6. RESULT AND DISCUSSION
Our similarity inference approach and industry standard Tanimoto measures conducted on the
same database and queries. Same evaluation method used for both. Result from the first
experiment is shown in Fig. 2, which shows the average number of similarly active compounds to
the target structures among the top 5% compounds retrieved. We found that our approach was
surpasses the industry standard Tanimoto measure in Antidepressants, Antiallergic/antiasthmatic,
AntiArrythmics and Antihypertensives activity groups tested. In Interacting on 5HT receptor,
Antiparkinsonians and Agents for Heart Failure activity groups our approach was found inferior to
the industry standard Tanimoto measures.
55 52
No of Active Compound in Top
50 47
44
45 41
40 36 35
35
30 28 27 28
26
5%
25
25 22
20 18
16
15
10
5
0
1 2 3 4 5 6 7
Activity Groups
Similarity Inference Industry Standard Tanimoto measure
FIGURE 2: Performance of Similarity Inference Network Compared to Performance of
Industry Standard Tanimoto Measure using BCI 2D bit string.
International Journal of Biometric and Bioinformatics, Volume (2) : Issue (1) 11
15. Ammar Abdo, Naomie Salim
40 37
No of Active Compound in
35 32
30
25 23 22
Top 5%
21 21
20 17
16 16
13 14 15 14 13
15
10
5
0
1 2 3 4 5 6 7
Activity Groups
Similarity Inference Industry standard Tanimoto Measure
FIGURE 3: Performance of Similarity Inference Network Compared to Performance of
Industry Standard Tanimoto Measure using Topological Indices.
Fig. 3 shows result from the second experiment. We found that our approach was surpasses the
industry standard Tanimoto measures in Interacting on 5HT receptor, Antidepressants and
AntiArrythmics activity groups tested. In Antiparkinsonians and Agents for Heart Failure activity
groups our approach was found inferior to the industry standard Tanimoto measures. In
Antiallergic/antiasthmatic and Antihypertensives activity groups, we found that both of the
approaches perform similarly.
50
No of Active Compound in
40
30
Top 5%
20
10
0
1 2 3 4 5 6 7
Activity Groups
Similarity Inference Using BCI
Similarity Inference Using Topological Indices
FIGURE 4: Performance of Similarity Inference Network Using BCI Compared to
Performance of Similarity Inference Network Using Topological Indices.
Fig. 4 shows the average number of similarly active compounds to the target structures among
the top 5% compounds retrieved. We found that our approach with bit-string descriptors from BCI
was performing better than when used with topological indices.
There are two distinct factors influence on the result produced by our approach. For 2D bit-string,
the no of common bits between compound and query (kjq), and the inverse fragment frequency
International Journal of Biometric and Bioinformatics, Volume (2) : Issue (1) 12
16. Ammar Abdo, Naomie Salim
(iff) of the fragment in the collection. For topological indices, the distance between descriptors
values of query and compound, and weight of query descriptor nodes (wi).
These factors constitute the weighting functions used in our approach. These weighting function
are intended to increase the influence of fragments and descriptors that are believed to be
important on quantifying the similarity. The basic ideas are that
• Many bits share by compound and query lead to increase the similarity score of this
compound
• Those fragments that occurs infrequently in the collection are more likely to be important than
frequent fragments and increase the similarity score of this compound.
• Slight distance between descriptor values lead to increase the similarity score of this
compound
7. CONSLUSION & FUTURE WORK
We have notice that the existing molecular similarity searching methods suffer from problems like
instability, unstandardize and poor results. The instability appears because no judgment can be
made about which best coefficients can be used for all biological activities. The similarity method
can start with little information, and as a general rule, the molecular similarity concept is most
often applied when knowledge of the system is sparse. This one of the advantage of molecular
similarity method but at the same time is disadvantage to these methods.
In this work we are proposing Bayesian inference networks for molecular similarity searching. We
have developed a novel approach for molecular similarity based on Bayesian inference networks,
which can resolve these problems. Our approach can comprise belief, weights and any other
evidences in the problem of molecular similarity. Overall results show the networks performed
slightly improvement than industry standard Tanimoto measures. We foresee that the result can
be much better when a better weighting function can be devised. Currently, we are working on
developing new weighting functions which include the frequency of each fragment in compound
to use in our similarity inference network.
8. REFERENCES
1. M. A. Johnson and G. M. Maggiora. “Concepts and Application of Molecular Similarity”, John
Wiley & Sons, New York (1990)
2. R. P. Sheridan and S. K. Kearsley. “Why do we need so many chemical similarity search
methods?”. Drug Discov. Today, 7, 903–911, 2002
3. M. A. Miller. “Chemical Database Techniques in Drug Discovery”. Nature Reviews Drug
Discov.,1, pp. 220-227, 2002
4. P. Willett. “Chemoinformatics: an application domain for information retrieval techniques”. In
Proceedings of the 27th Annual international ACM SIGIR Conference on Research and
Development in information Retrieval SIGIR '04. ACM, New York, NY, 393-393, 2004
5. P. Willett, J. M. Barnard and G. M. Downs. “Chemical similarity searching”. Journal of
Chemical Information and Computer Sciences, 38:983-996, 1998
6. P. M. Dean. “Molecular Similarity In Drug Design”. Blackie Academic & Professional, London,
1995
International Journal of Biometric and Bioinformatics, Volume (2) : Issue (1) 13
17. Ammar Abdo, Naomie Salim
7. R. E. Carhart, D. H. Smith and R. Venkataraghavan. “Atom pairs as molecular features in
structure-activity studies: definitions and applications”. Journal of Chemical Information and
Computer Science, 25:64-73, 1985
8. D. E. Patterson, R. D. Cramer, A. M. Ferguson, R. D. Clark and L. E. Weinberger.
“Neighborhood behavior: as useful concept for validation of molecular diversity descriptors”.
Journal of Medical Chemistry, 39:3060-3069, 1996
9. P. Willett, V. Winterman and D. Bawden. “Implementation of nearest neighbour searching in
an online chemical structure search system”. Journal of Chemical Information and Computer
Science, 26:36-41, 1986
10. T. R. Hagadone. “Molecular substructure similarity searching: efficient retrieval in two-
dimensional structure databases”. Journal of Chemical Information and Computer Science.
32:515-521, 1992
11. W. Fisanick, K. P. Cross and A. Rusinko. “Similarity searching on CAS Registry Substances.
1. Global molecular property and generic atom triangle geometric searching”. Journal of
Chemical Information and Computer Sciences, 32:664-674, 1992
12. D. Bawden. “Molecular dissimilarity in chemical information systems”. In Chemical Structures
Vol. 2: The International Language of Chemistry (W. A. Warr, ed.), Springer-Verlag,
Hiedelberg, pp. 383-388, 1993
13. M. S. Lajiness. “Dissimilarity-based compound selection techniques”. Perspectives in Drug
Discovery and Design, 7/8:65-84, 1997
14. E. J. Martin, J. M. Blaney, M. A. Siani, D. C. Spellmeyer, A. K. Wong and W. H. Moos.
“Measuring diversity: Experimental design of combinatorial libraries for drug discovery.
Journal of Medicinal Chemistry, 38:1431-1436, 1995
15. J. D. Holliday and P. Willett. “Definitions of "dissimilarity" for dissimilarity-based compound
selection”. Journal of Biomolecular Screening, 1:145-151, 1996
16. V. J. Gillet, P. Willett and J. Bradshaw. “The effectiveness of reactant pools for generating
structurally diverse combinatorial libraries”. Journal of Chemical Information and Computer
Science. 37:731-740, 1997
17. P. Willett. “Similarity-based virtual screening using 2D fingerprints”. Drug Discov. Today,
1046-1053, 2006
18. P. G. Dittmar, N. A. Farmer, W. Fisanick, R. C. Haines and J. Mockus. “The CAS online
search system. 1. General system design and selection, generation and use of search
screens”. Journal of Chemical Information and Computer Sciences, 23:93-102, 1983
19. Barnard Chemical Information Ltd., “Barnard Chemical Information Fingerprint Software
Documentation”. MAKEBITS version 3.3, p. 1-5, 1997
20. Barnard Chemical Information Ltd., “Barnard Chemical Information Fingerprint Software
Documentation”. MAKEFRAG version 3.3, Sheffield, p. 1, 1997
International Journal of Biometric and Bioinformatics, Volume (2) : Issue (1) 14
18. Ammar Abdo, Naomie Salim
21. J. L. Durant, B. A. Leland, D. R. Henry and J. G. Nourse. “MDL keys revisited”. 2nd Joint
Sheffield Conference on Chemoinformatics: Computational Tools For Lead Discovery,
University of Sheffield, Sheffield, 2001
22. J. L. Durant, B. A. Leland, D. R. Henry and J. G. Nourse. “Reoptimization of MDL keys for
use in drug discovery”. Journal of Chemical Information and Computer Science, 42:1273-
1280, 2002
23. Tripos Inc. UNITY Reference Guide version 4.1. Tripos, St. Louis, Missouri, 1999
24. C. A James, D. Weininger and J. Delany. “Daylight Theory Manual”
http://www.daylight.com/dayhtml/doc/theory/index.html
25. G. M. Downs and P. Willett. “Similarity searching in databases of chemical structures”. In: K.
B. Lipkowitz and D. B. Boyd (Eds.), Reviews in Computational Chemistry, VCH Publishers,
New York, Vol. 7, pp. 1-66, 1996
26. L. Hodes. “Clustering a large number of compounds. 1. Establishing the method on an initial
sample”. Journal of Chemical Information and Computer Science, 29:66-71, 1989
27. P. Willett and V. Winterman. “A comparison of some measures of intermolecular structural
similarity”. Quantitative Structure-Activity Relationships, 5, 18–25, 1986
28. P. Willett. “Algorithms for calculation of similarity in chemical structure databases”. In
Concepts and Application of Molecular Similarity, M. A. Johnson and G. M. Maggiora, Eds.,
John Wiley and Sons, New York. pp. 43-61, 1990
29. P. H. A. Sneath and R. R. Sokal. “Numerical Taxanomy”. Freeman, San Francisco, 1973
30. P. Willett. “Similarity And Clustering In Chemical Information Systems”, Research Studies
Press, Letchworth, (1987)
31. D. Ellis, J. Furner-Hines and P. Willett. “Measuring the degree of similarity between objects in
text retrieval systems”. Perspective in Information Management. 3:128-149, 1993
32. G. W. Adamson and J. A. Bush. “A method for the automatic classification of chemical
structures”. Information Storage and Retrieval, 9:561-568,1973
33. J. Pearl. “Probabilistic reasoning in intelligent systems: Networks of plausible inference”,
Morgan Kaufmann Publishers, (1988)
34. G. Salton and M. J. McGill. “Introduction to Modern Information Retrieval”, McGraw-Hill,
NewYork, (1983)
35. C. J. Van Rijsbergen. “Information Retrieval”, 2nd ed., University of Glasgow, 87-110 (1979)
36. H. Turtle. “Inference Networks for Document Retrieval”. PhD Thesis, University of
Massachusetts, 1990
37. H. Turtle and W. Croft. “A comparison of text retrieval models”. Comput. Journal, 35, 279-
290, 1992
International Journal of Biometric and Bioinformatics, Volume (2) : Issue (1) 15
19. Ammar Abdo, Naomie Salim
38. B. A. N. Ribeiro and R. Muntz. “A belief network model for IR”. In: Proceedings of the 19th
ACM SIGIR Conference, pp. 253–260,1996
39. S. K. M. Wong and Y. Y Yao. “On modeling information retrieval with probabilistic inference”.
ACM Transactions on Information Systems, Vol. 13, No. 1, pp. 38-68, 1995
40. H. Turtle and W. Croft. “Evaluation of an inference network-based retrieval model”. ACM
Transactions on Information Systems, 9:187-222, 1991
41. Barnard Chemical Information Ltd., “Barnard Chemical Information Fingerprint”.
http://www.bci.gb.com
42. J. Gasteiger and T. Engel. ”Chemoinformatics”, VCH-Wiley, New York, Vol. 1, pp. 3-5 (2003)
43. L. M. De Campos, J. M. Fernández and J. F. Huete. “The BNR model: foundations and
performance of a Bayesian network- based retrieval model”. Int. J. Approx. Reasoning, 3, pp.
265–285, 2003
44. Molecular Design Ltd., MDDR “MDL Drug Data Report Database”. http://www.mdli.com
45. Melano Chemoinformatics. “Dragon software”. http://www.talete.mi.it
46. N. Salim, J. Holliday and P. Willet. “Combination of fingerprint-based similarity coefficients
using data fusion”. J. Chem. Inf. Comput. Sci., 43, pp. 435-442, 2003
47. P.A. Bath, C. A. Morris and P. Willett. “Effect of standardisation of fragment-based measures
of structural similarity”. Journal of Chemometrics, 7, pp. 543, 1993.
48. N. Daut, R. Mohemad and N. Salim. “Finding Best Coefficients for Similarity Searching Using
Neural Network Algorithm”. International Conference in Artificial Intelligence in Engineering &
Technology (ICAIET), 2006.
49. Downs, G.M., Poirrette, A.R., Walsh, P. and Willett, P. “Evaluation of similarity searching
methods using activity and toxicity data”. In Chemical Structures Vol. 2: The International
Language of Chemistry (W. A. Warr, ed), Springer Verlag, Heidelberg, pp. 409-421, 1993
50. N. Salim and W. W. P. Godfrey. “Effectiveness of Probability Models for Compound Similarity
Searching”. Journal of Advancing Information Management Studies, 2(1): pp. 56-74, 2005.
International Journal of Biometric and Bioinformatics, Volume (2) : Issue (1) 16
